Research On Road Scene Segmentation Method Based On Convolutional Neural Network

Image semantic segmentation is an important research topic in computer vision.It is one of the key technologies for scene understanding.It has a wide range of applications in autonomous driving,medical image segmentation,wearable devices and many other fields.The traditional segmentation method is limited by the ability to extract image features and cannot meet the accuracy requirements of semantic segmentation tasks in complex scenes.With the development of deep learning,convolutional neural networks began to be applied to many fields of computer vision.The emergence of the full convolutional neural network makes the classification of the target in the deep learning to the pixel level,which greatly improves the accuracy and speed of the semantic segmentation task.Image semantic segmentation applied to road scenes can be divided into two types: segmentation of different objects in the image and finer segmentation of different targets of the same class of objects.The problem of road scene segmentation studied in this paper is to realize multiple segmentation of different types of objects in road scenes,which can be divided into 19 categories: roads,sidewalks,trees,buildings,pedestrians,and automobiles.It is required to speed up the detection speed of the network as much as possible without reducing the accuracy of detection.Aiming at the above problems,this thesis conducts in-depth research on various neural network structures and semantic segmentation models,and proposes two semantic segmentation models based on encoder-decoder structure.These two network structures are improvements to the classical semantic segmentation network UNet.The main work and contributions are as follows:1.The linear feature interpolation is used to upsample the image feature matrix to reduce the training difficulty,and the weight of the low-level features and the upsampling result in the jump connection is balanced by introducing a 1x1 convolution structure.2.The introduction of deep separable convolution structure reduces the complexity of the model,improves the computational speed of semantic segmentation,and makes the semantic segmentation network easier to converge on small batch training sets.3.By introducing the bottleneck structure,the network can better preserve lowlevel feature information when down sampling.4.A hole convolution structure is introduced to reduce the number of down sampling of the network.Finally,the experimental results show that the improved model is effective and feasible for the performance improvement of road scene semantic segmentation tasks.